V预测模型LoRA训练指南
V-Prediction SDXL Model LoRA Training Guide
此教程适用于基于sd-scripts的LoRA训练器。
This tutorial is intended for LoRA trainers based on sd-scripts.
目前kohya-ss/sd-scripts训练器仅在dev分支对在v预测模型上训练LoRA进行了支持，使用主分支会导致花图。
Currently, the kohya-ss/sd-scripts trainer only supports training LoRA on v-prediction models in the dev branch. Using the main branch may result in corrupted images.
下面以秋叶LoRA训练包为例进行演示。
The following example uses the Aki LoRA training package.
在存放有训练包的文件夹下按shift+右键，点击“在终端中打开”。
In the folder containing the training package, press Shift + Right-click and select "Open in Terminal.“
在弹出的命令行窗口内输入以下指令，clone dev分支的训练器到本地：
In the command line window that appears, enter the following command to clone the dev branch of the trainer locally:git clone -b dev https://github.com/kohya-ss/sd-scripts.git
（对于中国大陆用户，如遇网络问题，请自行打开VPN/加速工具/使用镜像）
进入秋叶包内，将原来的训练器文件夹改名做备份。不建议删除原来的，因为dev分支可能有bug。
Go into the Autumn Leaf package and rename the original trainer folder to create a backup. It’s not recommended to delete the original, as the dev branch may contain bugs.
将新clone下来的sd-scripts拖到秋叶包内。
Drag the newly clonedsd-scriptsinto the Aki package.
启动WebUI脚本，打开训练webUI。
Start the WebUI script and open the training webUI.在高级设置处添加如下参数：
In the advanced settings, add the following parameter:v_parameterization = true zero_terminal_snr = true
其余设置按照自己平时训练配置即可。
The rest of the settings can follow your usual training configuration.当训练时弹出如下警告，说明已开启v预测模式。
When the following warning appears during training, it indicates that v-prediction mode is enabled.
训练成果 training result (nyalia lora on noob-vpred0.5)：
以上即为全部教程。
That concludes the tutorial.
如果你使用命令行训练器/linux，那么仅需在切换sd-scripts为dev分支后，在训练时为训练参数添加--v_parameterization true
If you are a command-line trainer/Linux user, simply switch sd-scripts to the dev branch, then add --v_parameterization true to the training parameters during training.
如遇bug，请自行把sd-script文件夹改名改回去。
If you encounter any bugs, rename the sd-scripts folder back to its original name.